CN118214676B - Network fault control method based on parallel network large model digital expert - Google Patents
Network fault control method based on parallel network large model digital expert Download PDFInfo
- Publication number
- CN118214676B CN118214676B CN202410320595.2A CN202410320595A CN118214676B CN 118214676 B CN118214676 B CN 118214676B CN 202410320595 A CN202410320595 A CN 202410320595A CN 118214676 B CN118214676 B CN 118214676B
- Authority
- CN
- China
- Prior art keywords
- network
- model
- learning
- parallel
- module
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 60
- 230000003993 interaction Effects 0.000 claims abstract description 35
- 238000012549 training Methods 0.000 claims abstract description 27
- 230000002452 interceptive effect Effects 0.000 claims abstract description 22
- 238000005516 engineering process Methods 0.000 claims abstract description 19
- 238000011217 control strategy Methods 0.000 claims abstract description 18
- 238000003058 natural language processing Methods 0.000 claims abstract description 14
- 238000012545 processing Methods 0.000 claims abstract description 10
- 238000007726 management method Methods 0.000 claims description 32
- 238000002474 experimental method Methods 0.000 claims description 27
- 238000005457 optimization Methods 0.000 claims description 27
- 230000008569 process Effects 0.000 claims description 19
- 238000011156 evaluation Methods 0.000 claims description 17
- 238000004458 analytical method Methods 0.000 claims description 16
- 230000006399 behavior Effects 0.000 claims description 16
- 238000004422 calculation algorithm Methods 0.000 claims description 16
- 238000005094 computer simulation Methods 0.000 claims description 13
- 238000013461 design Methods 0.000 claims description 12
- 238000004364 calculation method Methods 0.000 claims description 10
- 230000009471 action Effects 0.000 claims description 7
- 238000013401 experimental design Methods 0.000 claims description 7
- 238000013135 deep learning Methods 0.000 claims description 5
- 238000010801 machine learning Methods 0.000 claims description 4
- 238000013528 artificial neural network Methods 0.000 claims description 3
- 238000004140 cleaning Methods 0.000 claims description 3
- 238000007418 data mining Methods 0.000 claims description 3
- 230000002068 genetic effect Effects 0.000 claims description 3
- 238000002372 labelling Methods 0.000 claims description 3
- 230000002787 reinforcement Effects 0.000 claims description 3
- 238000012544 monitoring process Methods 0.000 claims description 2
- 230000008901 benefit Effects 0.000 abstract description 3
- 239000003795 chemical substances by application Substances 0.000 description 9
- 238000010276 construction Methods 0.000 description 4
- 238000004088 simulation Methods 0.000 description 4
- 238000009411 base construction Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 230000002159 abnormal effect Effects 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- 102100033814 Alanine aminotransferase 2 Human genes 0.000 description 1
- 101710096000 Alanine aminotransferase 2 Proteins 0.000 description 1
- 239000008186 active pharmaceutical agent Substances 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000003449 preventive effect Effects 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
- H04L41/149—Network analysis or design for prediction of maintenance
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/06—Management of faults, events, alarms or notifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/08—Configuration management of networks or network elements
- H04L41/0876—Aspects of the degree of configuration automation
- H04L41/0886—Fully automatic configuration
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/08—Configuration management of networks or network elements
- H04L41/0894—Policy-based network configuration management
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
- H04L41/145—Network analysis or design involving simulating, designing, planning or modelling of a network
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/16—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks using machine learning or artificial intelligence
Landscapes
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Automation & Control Theory (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Databases & Information Systems (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a network fault management and control method based on a parallel network large model digital expert. The method comprises the steps of extracting characteristics and rules of a network from data of an actual network, constructing behavior and interaction rules of an artificial network intelligent agent, generating a network scene, designing an experimental scheme aiming at a specific fault network scene, training and evaluating control strategies under the corresponding scene, selecting an optimal control strategy, integrating the fault network scene and the optimal control strategy into a large model digital expert model, constructing an interactive question-answer corpus knowledge base, understanding and processing natural language texts of the interactive question-answer corpus knowledge base by utilizing a natural language processing technology, constructing a fine-tuned parallel network field model, and realizing dynamic selection of a parallel network through a human-computer interaction system. The method utilizes the advantages of parallel networks and combines a large-model digital expert model to realize an effective network fault management and control method.
Description
Technical Field
The invention belongs to the crossing field of computer science and network technology, and particularly relates to a network fault management and control method based on a parallel network large model digital expert.
Background
With popularization of the internet and rapid development of informatization technology, networks are widely used in various fields, such as cloud computing, internet of things, industrial internet, and the like. However, during the operation of the network, various faults and abnormal conditions, such as network congestion, node failure, malicious attack, etc., often occur, and these faults and abnormal conditions may have serious influence on the stability and security of the network. Therefore, the management and control of various fault network scenes in the network are necessary means for guaranteeing the stability and safety of the network.
At present, the prior art is a network management and control method based on a parallel network architecture, and the method realizes the optimal management and control of an actual system, the experiment and evaluation of related behaviors and decisions, and the study and training of related personnel and systems through virtual-real interaction and parallel execution between an artificial network and an actual network. Specifically, the method comprises the following steps: firstly, constructing an artificial network, wherein the artificial network and an actual network have the same network structure and node attribute; then, transmitting the flow data and the node state data in the actual network to the artificial network through the virtual-real interaction technology; then, simulating the operation process of an actual network in a manual network, and carrying out feature extraction and classification on flow data and node state data by using a large model technology; and finally, carrying out optimal management and control on the actual network according to the prediction result, thereby realizing the optimal management and control on the actual system.
However, the management and control of the network by the above method still has the following problems: first, accuracy and efficiency problems: the simulation accuracy and efficiency of a human network or parallel system may be affected by many factors, such as model design, computational power, data quality, etc. If the simulation accuracy and efficiency are not high, the management and control effects of the actual network are affected. Second, real-time and robustness: real-time and robustness of an artificial network or parallel system may also be affected by many factors, such as virtual-real interaction mechanism, system stability, network dynamic changes, etc. If the real-time performance and the robustness are not good, the management and control effects of the actual network are affected. Third, complexity and scalability: the complexity and scalability of manual networks or parallel systems is also a technical problem. In practical applications, the design and implementation of a manual network or parallel system can be very complex and difficult to expand and accommodate new situations and requirements. Fourth, interactivity and controllability: the interactivity and controllability of a manual network or a parallel system with the actual network is also a critical issue. If the artificial network or the parallel system cannot effectively interact and control with the actual network, the management and control effects thereof are limited.
Disclosure of Invention
In order to solve the technical problems, the invention provides a network fault management and control method based on a parallel network large model digital expert, which utilizes the advantages of a parallel network, by combining a large-model digital expert model, an effective network fault management and control method is realized, faults are quickly repaired, and the reliability and stability of the system are improved.
In order to achieve the above purpose, the technical scheme adopted by the invention is as follows:
a network fault management and control method based on a parallel network large model digital expert comprises the following steps:
Step 1: building a parallel network architecture, and realizing optimal management and control of an actual network, experiments and evaluation of related behaviors and decisions, and study and training of related personnel and systems through virtual-real interaction and parallel execution between an artificial network and the actual network;
Step 2: an experimental design module is constructed, the combination of various network agents is designed through an artificial network, a plurality of fault network scenes are generated based on preset rules, and a corresponding experimental scheme is generated in an auxiliary mode through intelligent planning and expert system methods;
step 3: constructing a dynamic simulation module, executing an experimental scheme generated by the experimental design module, and evolving complete artificial network data;
step 4: constructing a learning optimization module, and training and learning the artificial network data generated by the dynamic simulation module by combining with the actual network data to obtain control strategies of experimental schemes corresponding to a plurality of fault network scenes;
Step 5: an analysis evaluation module is constructed, the process and result data of the learning optimization module are evaluated and analyzed, and an optimal control strategy conforming to an experimental target is judged;
Step6: constructing a data center module, and storing the respective corresponding optimal control strategies under a plurality of fault network scenes;
Step 7: various information generated by an experiment design module, a dynamic simulation module, a learning optimization module, an analysis evaluation module and a data center module is integrated into a large-model digital expert intelligent interaction system, the large-model digital expert intelligent interaction system converts the information into characters and constructs an interactive question-answer corpus knowledge base, natural language processing technology is utilized to understand and process natural language texts of the interactive question-answer corpus knowledge base, and a finely-tuned parallel network model is built regularly;
And 8, constructing a man-machine interaction system, and dynamically selecting a local network model and/or a fine-tuned latest parallel network model.
Further, the step 2 includes acquiring an actual network signal and a user signal through a physical signal sensor and a social signal sensor; learning the obtained actual network signals and the user signals by means of machine learning and data mining to obtain parameters of an actual network; the parameters of the actual network are analyzed through natural language processing and parallel learning technologies, the requirements and optimization targets of users on the actual network are obtained, the behaviors and interaction rules of the artificial network intelligent agent are automatically generated, the corresponding fault network scene is further generated, and the corresponding experimental scheme is generated in an auxiliary mode through intelligent planning and expert system methods.
Further, the step 3 includes generating a certain number of agents based on the fault network scene and the interaction rule generated by the experiment design module, simulating their behaviors and interactions, collecting experimental process data, and simultaneously marking and/or processing the data according to the specified targets.
Further, the step 4 includes, combining the data to be marked and/or processed, and obtaining different control strategies under various fault network scenes by using a learning and training method; the learning and training method comprises at least one algorithm of fuzzy control, dynamic programming, an artificial neural network, a genetic algorithm, an ant colony algorithm, deep learning and reinforcement learning, and the learning optimization module has self-learning and adjustment capability and can select and optimize the algorithm used in the calculation experiment.
Further, the step 5 includes that the analysis and evaluation module is further used for checking the learning optimization module to see whether the experiment process generates errors or not and analyzing and correcting the errors.
Further, the step 6 includes the data center module providing fast query of the optimal control strategy of the corresponding fault network scenario based on the parameters of the actual network.
Further, in the step 7, constructing an interactive question-answer corpus knowledge base includes constructing a corpus, and constructing a knowledge base, wherein constructing the corpus includes collection and cleaning processing of the corpus, basic information statistics of the corpus and corpus analysis of the interactive question-answer; the knowledge base is constructed to include topic labels, user behavior labels and context labels of question and answer sentences.
Further, in the step 7, the periodically constructing the trimmed parallel network model by using the natural language processing technology includes:
step 7.1, extracting natural language commonality characteristics on an interactive question-answer corpus knowledge base by the GPT model, and pre-training the GPT model in a self-supervision learning mode;
step 7.2, primarily introducing real person value preference based on answers of user behavior labeling preference, and monitoring and fine-tuning a GPT model according to a manual example;
step 7.3, based on the sorting of real people on the output data of the fine-tuning GPT model, supervising and training to obtain a reward model, so that the reward model learns the value preference of the real people;
And 7.4, optimizing the fine-tuning GPT model, namely the parallel network model, by adopting a near-end strategy to optimize PPO algorithm closed-loop feedback based on the reward model.
Further, the step 8 includes that the man-machine interaction system speech recognizes the input of the user, and carries out natural language understanding on the input to recognize the intention and the slot related to the task; then, tracking dialogue state, combining history dialogue state, and updating to obtain current dialogue state based on the identified intention and slot position; and generating an optimal action based on the current dialogue state, and generating a reply according to the optimal action to be executed and returning the reply to the user.
The invention has the beneficial effects that:
1. By combining a parallel network and a large model digital expert system, network faults can be detected and identified more quickly, the fault discovery time is shortened, the influence of faults on the system performance is reduced, and the recovery speed and the processing capacity of the system are improved; the intelligent fault management and control are realized, the network configuration and management strategy are automatically adjusted, and the intelligent degree and the adaptability of the system are improved.
2. The large-model digital expert system can predict potential faults and perform preventive maintenance based on historical data and current network states, reduces the probability of system shutdown and faults, and improves the reliability and stability of the system.
3. By combining the parallel network and the large-model digital expert system, more intelligent fault management and control can be realized, network configuration and management strategies can be automatically adjusted, and the intelligent degree and the adaptability of the system are improved.
Drawings
FIG. 1 is a flow chart of the application of the parallel network large model digital expert-based network fault management and control method of the invention;
FIG. 2 is a diagram of a parallel network architecture according to an embodiment of the present invention;
FIG. 3 is a flow chart of a large model digital expert intelligent interactive system according to an embodiment of the invention;
FIG. 4 is a flowchart of an interactive question-answer corpus knowledge base construction according to an embodiment of the present invention;
FIG. 5 is a diagram of a pre-trained language model according to an embodiment of the present invention;
FIG. 6 is a flowchart of an implementation of constructing a trim ChatGPT model in accordance with an embodiment of the present invention;
FIG. 7 is a flow chart illustrating a task session process according to an embodiment of the present invention;
fig. 8 is a schematic diagram of an open field dialog framework based on deep learning according to an embodiment of the present invention.
Detailed Description
The invention will be further described with reference to the accompanying drawings and examples.
In the invention, various fault scenes in the network construct a parallel network, and the core is a calculation experiment. The main steps of the calculation experiment are to extract the characteristics and rules of the network from the data of the real network based on machine learning, statistical analysis and other technologies. Then, based on the extracted characteristics, constructing behavior and interaction rules of an artificial network intelligent agent and generating a network scene to realize simulation of a real network, on the basis, designing a specific experimental scheme aiming at specific targets and applications, further training and evaluating a control strategy under the corresponding scene based on complete artificial network data, taking the control strategy which is most in line with the experimental targets as an optimal control strategy of the network, and then inputting various output information of an experimental design module, a dynamic simulation module, a learning optimization module, an analysis evaluation module and a data center module into a large-model digital expert intelligent interaction system to be converted into characters and construct an interactive question-answer corpus knowledge base, and carrying out man-machine webpage version interaction through the knowledge base to dynamically expand information; and understanding and processing natural language texts of the interactive question-answer corpus knowledge base by using a natural language processing technology, and periodically constructing a fine-tuned parallel network field model.
As shown in fig. 1, a flow chart of the network fault control method based on the parallel network large model digitization expert is provided, and the method flow is implemented by the following five modules: the experimental design module, the dynamic simulation module, the learning optimization module, the analysis evaluation module and the data center module interact and cooperate to realize the specific implementation processes of analysis, prediction and guidance of the complex network system as follows:
Step 1: the parallel network architecture is built, the artificial network architecture in the parallel network adopts a method for creating virtual network containers which are logically isolated from each other but can share the same underlying physical network, so that the underlying hardware is controlled through the programmability of a software platform in the centralized controller, and flexible on-demand allocation of network resources is realized. In the artificial network, the network equipment is only responsible for simple data forwarding and can adopt general hardware; the original operating system in charge of control is extracted into an independent network operating system in charge of adapting different service characteristics, and the communication between the network operating system and the service characteristics and between the network operating system and the hardware devices can be realized through programming. Flexible reconfiguration of network functions is achieved through the open software defined APIs, and expansion capacity and flexibility of the network are improved. The control application need only focus on its own logic and not on the underlying more detailed implementation. While a logically centralized control plane can control a plurality of forwarding plane devices, i.e. the entire physical network. Therefore, a global network state view can be obtained, and the optimization of the network is realized according to the global network state, so that various requirements of users are met. Fig. 2 shows a parallel network architecture, and through virtual-real interaction and parallel execution between a manual network and an actual network, optimal management and control of an actual system, experiments and evaluation of related behaviors and decisions, and learning and training of related personnel and systems are realized.
Step 2: the experimental design module designs the combination of various network agents through a manual network, namely a computing laboratory, and generates a plurality of fault network scenes based on a certain rule, wherein the fault network scenes comprise discrete events such as packet loss, delay and the like which occur randomly in the network. Specifically, it is to acquire an actual network signal and a user signal through a physical signal sensor and a social signal sensor. Learning the network by means of machine learning, data mining and other technologies so as to obtain parameters of an actual network; through knowledge automation technologies such as natural language processing, parallel learning and the like, the system can analyze the demands of users on a network, so as to understand network optimization tasks, and can automatically generate behaviors and interaction rules of artificial network agents based on the analyzed experimental targets and schemes, so as to generate corresponding network scenes. And an experimental scheme corresponding to the network fault scene is generated in an auxiliary way through methods such as intelligent planning, expert systems and the like.
Step 3: the dynamic simulation module is based on a traditional network simulation platform, can execute network experiments based on intelligent agents, and evolves complete artificial network data. The intelligent experiment system mainly has the functions of generating a certain number of intelligent agents based on network scenes and rules generated by an experiment design module, simulating the behaviors and interactions of the intelligent agents, collecting experimental process data, and marking or processing the data according to specified targets. And then, different strategies under various fault scenes are obtained by utilizing a learning and training method, and different schemes are evaluated based on a certain strategy, so that an optimal scheme aiming at different fault network scenes is obtained.
Step 4: the learning optimization module analyzes and learns the big data of the artificial network generated by dynamic simulation in combination with the actual network data, thereby acquiring specific knowledge of the actual network and the artificial network and generating an optimized parameter adjusting scheme of the actual network and the artificial network. The learning optimization module comprises control algorithms such as fuzzy control, dynamic programming, an artificial neural network, a genetic algorithm, an ant colony algorithm and the like, and also comprises artificial intelligent algorithms such as deep learning, reinforcement learning and the like, and can optimize network parameters based on a large number of calculation experiments. Meanwhile, the learning optimization module also has certain self-learning and adjusting capacity, and can flexibly select and optimize an algorithm used for calculation experiments.
Step 5: the analysis evaluation module evaluates and analyzes the process and result data of the learning optimization module so as to judge an optimal control strategy which is most in line with an experimental target, and checks the experimental process to see whether the experimental process generates errors or not and analyze and correct the errors.
Step 6: based on a large number of scene experiments and training evaluations, the computing experiment platform can generate a large number of optimization schemes and control strategies for different scenes. In order to facilitate quick presentation of the corresponding solutions in the face of new network conditions, the data center module will store the optimal solutions in various network scenarios and can quickly query the solutions for the corresponding scenarios based on parameters of the actual network.
Step 7: the large model digital expert is integrated into network fault management and control, so that the calculation experiment of the network fault management and control has thinking chain capability based on an intelligent interaction system, namely, logical deduction from phenomenon to principle is realized, logical reasoning among concepts can be completed, and stronger knowledge understanding capability and contextual learning capability are provided. The general flow of the intelligent interactive system based on large model digitized experts is substantially similar to that shown in fig. 3. Various information generated by an experiment design module, a dynamic simulation module, a learning optimization module, an analysis evaluation module and a data center module is integrated into a large-model digital expert intelligent interaction system, the large-model digital expert intelligent interaction system converts the information into characters and constructs an interactive question-answer corpus knowledge base, natural language processing technology is utilized to understand and process natural language texts of the interactive question-answer corpus knowledge base, and a finely-tuned parallel network model is built regularly;
Wherein, constructing the corpus comprises: and the information generated by the experiment design module, the dynamic simulation module, the learning optimization module and the analysis evaluation module in the calculation experiment for constructing the parallel network for various fault scenes of the cluster network is converted into characters by the intelligent interaction system. The method comprises the steps of constructing an interactive question-answer corpus knowledge base by an experiment design module, a dynamic simulation module, a learning optimization module and an analysis evaluation module in a calculation experiment of a parallel network for various fault scenes of a cluster network. As shown in fig. 4, the construction process mainly includes corpus construction and knowledge base construction, and the corpus construction step includes corpus collection and cleaning treatment, corpus basic information statistics and interactive question-answering corpus analysis; the knowledge base construction mainly comprises topic labels of question-answer sentences, user behavior labels and context relation labels.
Natural language processing includes: the calculation experiment of large model digital expert based on intelligent interactive system for cluster network fault management and control mainly uses natural language processing technology. Natural language processing is a process of understanding and processing natural language text by a computer and extracting text semantics. Only on the basis of understanding the text content, the intelligent interaction system can perform the next processing and operation according to the text semantics. Therefore, the natural language processing technology is an important link of the intelligent interaction technology. As shown in FIG. 5, as a pre-training language model, the main idea of the pre-training model is to learn general semantic representation by pre-training on large-scale text corpus data, and then transmit knowledge learned in the pre-training stage to a downstream task, so that dependence of the downstream task on a large amount of data is reduced, the convergence speed of the model is increased, and the generalization capability of the model on the downstream task is improved. The pre-training model represented by GPT-2 and GPT-3 explores a downstream knowledge migration mode without fine adjustment, and GPT-3 performs pre-training on a very large scale unsupervised corpus (45T) through a trans former model with very large parameters (1750 hundred million parameters), and learns natural language unified representation of various tasks by taking a language model as a target, so that knowledge capability of unsupervised reasoning on downstream tasks is obtained. Specifically, GPT-3 can be applied to downstream tasks in a small sample or zero sample learning mode, fine adjustment on the downstream task corpus is not needed, and only a small number of sample prompts are given to reasoning the downstream tasks. ChatGPT benefit from its successful introduction of a human value preference. Based on this, unlike other pre-training models, the human language habit is introduced into the model by adopting the manner of ChatGPT to RLHF, and the basic flow of ChatGPT implementation is shown in fig. 6 and can be roughly divided into the following 4 steps: step 0: the GPT is pre-trained. Based on a large-scale corpus, the GPT model is pre-trained in a self-supervised learning mode. And the GPT is enabled to extract natural language commonality characteristics on a large-scale corpus. Step 1: and (5) supervising and fine-tuning the GPT. The answer based on the real person labeling preference is initially introduced into the real person value preference, and the GPT is finely adjusted according to the manual example supervision. Step 2: and (5) designing a reward model. Based on the data of the model output ordering of the true man, the supervision training obtains the rewarding model, so that the rewarding model learns the value preference of the true man. Step 3: RL feedback optimizes GPT. And (3) performing closed-loop feedback optimization supervision on the trimmed GPT by adopting a near-end strategy optimization (Proximal policy optimization, PPO) algorithm based on the reward model to obtain ChatGPT. GPT after step 0 does not necessarily perform well on specific tasks, but has considerable potential to achieve excellent performance in multiple tasks through fine tuning or context learning modes. While steps 1 through 3 are key steps of ChatGPT, these steps successfully introduce human factors into GPT.
And 8, constructing a man-machine interaction system, and dynamically selecting a local network model and/or a fine-tuned latest parallel network model.
Man-machine conversation techniques can be categorized into task-type conversations and open-domain conversations. Task-type conversations are a form of man-machine conversations that help users to accomplish efficiently. The process flow of a task-based dialog as used herein can be generally represented by fig. 7. For user input (Query), natural language understanding (Natural Language Understanding, NLU) is first performed after speech recognition, i.e., task related intent and slots are identified. NLU can be regarded as a specific class of tasks that text content understands, and intent and slot will typically be predefined according to the specific task. Dialog state tracking (Dialog STATE TRACKER) is then performed, and in combination with the historical Dialog state, a new Dialog state is updated based on the identified intent and slot; the Dialog Policy (Dialog) module may then generate an optimal action based on the current Dialog state, such as asking the user for some missing slot information, etc. Finally, the natural language generation (Natural Language Generation, NLG) module returns a Response to the user based on the action to be performed, converting the system generated dialog action into a natural language Response that the user can understand.
In addition, the invention can also adopt other man-machine interaction designs, and the depiction of the open field dialogue system based on deep learning is shown in fig. 8. In addition to the single-round dialogue system and the multi-round dialogue system according to whether the history dialogue information is considered by the input, the method can be divided into a search formula, a generation formula and a method of combining search and generation according to the construction method: the user is driven to conduct continuous multi-round conversations by giving reasonable replies of semantic association to the user Query. Early open domain conversations were based on rule-based approaches such as ELIZA, PARRY, ALICE, etc.; rule-based methods require much manual involvement, can generally be summarized only for a given domain, and are difficult to quickly expand to more domains. Firstly, performing non-supervision generation training on a corpus irrelevant to a dialogue task, and then performing fine adjustment on the dialogue corpus; as the non-supervision corpus is generally larger in scale, the priori knowledge of the language sequence can be well learned, and the method can help to generate replies with longer length and more specific semantics.
In summary, the invention realizes more intelligent fault management and control by combining the parallel network and the large model digital expert system, automatically adjusts network configuration and management strategy, and improves the intelligent degree and the adaptability of the system.
The foregoing description of the embodiments has been provided for the purpose of illustrating the general principles of the invention, and is not meant to limit the invention thereto, but to limit the invention thereto, and any modifications, equivalents, improvements and equivalents thereof may be made without departing from the spirit and principles of the invention.
Claims (8)
1. The network fault management and control method based on the parallel network large model digital expert is characterized by comprising the following steps:
Step 1: building a parallel network architecture, and realizing optimal management and control of an actual network, experiments and evaluation of related behaviors and decisions, and study and training of related personnel and systems through virtual-real interaction and parallel execution between an artificial network and the actual network;
Step 2: an experimental design module is constructed, the combination of various network agents is designed through an artificial network, a plurality of fault network scenes are generated based on preset rules, and a corresponding experimental scheme is generated in an auxiliary mode through intelligent planning and expert system methods;
step 3: constructing a dynamic simulation module, executing an experimental scheme generated by the experimental design module, and evolving complete artificial network data;
step 4: constructing a learning optimization module, and training and learning the artificial network data generated by the dynamic simulation module by combining with the actual network data to obtain control strategies of experimental schemes corresponding to a plurality of fault network scenes;
Step 5: an analysis evaluation module is constructed, the process and result data of the learning optimization module are evaluated and analyzed, and an optimal control strategy conforming to an experimental target is judged;
Step6: constructing a data center module, and storing the respective corresponding optimal control strategies under a plurality of fault network scenes;
Step 7: various information generated by an experiment design module, a dynamic simulation module, a learning optimization module, an analysis evaluation module and a data center module is integrated into a large-model digital expert intelligent interaction system, the large-model digital expert intelligent interaction system converts the information into characters and constructs an interactive question-answer corpus knowledge base, natural language processing technology is utilized to understand and process natural language texts of the interactive question-answer corpus knowledge base, and a finely-tuned parallel network model is built regularly;
step 8, constructing a man-machine interaction system, and dynamically selecting a local network model and/or a fine-tuned latest parallel network model;
in the step 7, the periodically constructing the trimmed parallel network model by using the natural language processing technology includes:
step 7.1, extracting natural language commonality characteristics on an interactive question-answer corpus knowledge base by the GPT model, and pre-training the GPT model in a self-supervision learning mode;
step 7.2, primarily introducing real person value preference based on answers of user behavior labeling preference, and monitoring and fine-tuning a GPT model according to a manual example;
step 7.3, based on the sorting of real people on the output data of the fine-tuning GPT model, supervising and training to obtain a reward model, so that the reward model learns the value preference of the real people;
and 7.4, optimizing the fine-tuning GPT model, namely the parallel network model, by adopting a near-end strategy to optimize PPO algorithm closed-loop feedback based on the reward model.
2. The network fault management and control method based on the parallel network large model digital expert according to claim 1, wherein the step 2 comprises obtaining actual network signals and user signals through a physical signal sensor and a social signal sensor; learning the obtained actual network signals and the user signals by means of machine learning and data mining to obtain parameters of an actual network; the parameters of the actual network are analyzed through natural language processing and parallel learning technologies, the requirements and optimization targets of users on the actual network are obtained, the behaviors and interaction rules of the artificial network intelligent agent are automatically generated, the corresponding fault network scene is further generated, and the corresponding experimental scheme is generated in an auxiliary mode through intelligent planning and expert system methods.
3. The method for controlling network faults based on the parallel network large model digital expert according to claim 1, wherein the step 3 comprises generating a certain number of agents based on fault network scenes and interaction rules generated by an experiment design module, simulating the behaviors and interactions of the agents, collecting experimental process data, and simultaneously marking and/or processing the data according to specified targets.
4. The network fault management and control method based on the parallel network large model digital expert as claimed in claim 1, wherein the step 4 comprises the steps of obtaining different control strategies under various fault network scenes by utilizing learning and training methods in combination with data to be marked and/or processed; the learning and training method comprises at least one algorithm of fuzzy control, dynamic programming, an artificial neural network, a genetic algorithm, an ant colony algorithm, deep learning and reinforcement learning, and the learning optimization module has self-learning and adjustment capability and can select and optimize the algorithm used in the calculation experiment.
5. The parallel network large model digital expert based network fault management method according to claim 1, wherein the step 5 comprises the analysis and evaluation module further used for checking the learning optimization module, checking whether the experimental process generates errors and analyzing and correcting the errors.
6. The parallel network large model digital expert based network fault management and control method according to claim 1, wherein the step 6 comprises the step that the data center module provides fast query of the optimal control strategy of the corresponding fault network scene based on the parameters of the actual network.
7. The method for managing and controlling network faults based on the parallel network large model digital expert according to claim 1, wherein in the step 7, constructing an interactive question-answer corpus knowledge base comprises constructing a corpus and constructing a knowledge base, wherein constructing the corpus comprises collection and cleaning processing of the corpus, basic information statistics of the corpus and corpus analysis of the interactive question-answer; the knowledge base is constructed to include topic labels, user behavior labels and context labels of question and answer sentences.
8. The parallel network large model digital expert based network fault management and control method according to claim 1, wherein the step 8 comprises the steps of the man-machine interaction system voice recognizing the input of the user and performing natural language understanding on the input to recognize the task related intention and the slot position; then, tracking dialogue state, combining history dialogue state, and updating to obtain current dialogue state based on the identified intention and slot position; and generating an optimal action based on the current dialogue state, and generating a reply according to the optimal action to be executed and returning the reply to the user.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410320595.2A CN118214676B (en) | 2024-03-20 | 2024-03-20 | Network fault control method based on parallel network large model digital expert |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410320595.2A CN118214676B (en) | 2024-03-20 | 2024-03-20 | Network fault control method based on parallel network large model digital expert |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN118214676A CN118214676A (en) | 2024-06-18 |
| CN118214676B true CN118214676B (en) | 2024-08-30 |
Family
ID=91447114
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410320595.2A Active CN118214676B (en) | 2024-03-20 | 2024-03-20 | Network fault control method based on parallel network large model digital expert |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN118214676B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN119398106A (en) * | 2024-08-27 | 2025-02-07 | 国网山东省电力公司平邑县供电公司 | A kind of power distribution network fault processing equipment |
| CN119862470B (en) * | 2025-03-21 | 2025-05-27 | 浙江英集动力科技有限公司 | A heating intelligent decision-making system based on artificial intelligence large model |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114254649A (en) * | 2021-12-15 | 2022-03-29 | 科大讯飞股份有限公司 | A language model training method, device, storage medium and device |
| CN115186473A (en) * | 2022-07-07 | 2022-10-14 | 北京交通大学 | Scene perception modeling and verifying method based on parallel intelligence |
Family Cites Families (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2000209202A (en) * | 1999-01-14 | 2000-07-28 | Nec Corp | Network fault management system |
| US12026612B2 (en) * | 2017-06-02 | 2024-07-02 | Google Llc | Optimization of parameter values for machine-learned models |
| US11392830B2 (en) * | 2018-04-13 | 2022-07-19 | The Regents Of The University Of California | Devices and methods employing optical-based machine learning using diffractive deep neural networks |
| WO2020021319A1 (en) * | 2018-07-27 | 2020-01-30 | Yogesh Chunilal Rathod | Augmented reality scanning of real world object or enter into geofence to display virtual objects and displaying real world activities in virtual world having corresponding real world geography |
| US11148671B2 (en) * | 2019-09-06 | 2021-10-19 | University Of Central Florida Research Foundation, Inc. | Autonomous systems human controller simulation |
| CN114070710A (en) * | 2020-09-22 | 2022-02-18 | 北京市天元网络技术股份有限公司 | Communication network fault analysis method and device based on digital twin |
| CN112347795B (en) * | 2020-10-04 | 2024-07-02 | 北京交通大学 | Machine translation quality assessment method, device, equipment and medium |
| CN115529256A (en) * | 2021-06-24 | 2022-12-27 | 中兴通讯股份有限公司 | Obstacle avoidance method, system, electronic device and computer-readable storage medium |
| CN115562042A (en) * | 2022-11-07 | 2023-01-03 | 广西大学 | Power generation control method for predictive control of distributed parallel multilayer model |
| CN115730872A (en) * | 2022-11-23 | 2023-03-03 | 交控科技股份有限公司 | Urban rail transit parallel deduction system and method |
| CN116935004A (en) * | 2023-02-07 | 2023-10-24 | 太原理工大学 | A method for intelligent operation of multi-type human-machine collaboration in fully mechanized mining working faces |
| CN116740610A (en) * | 2023-06-15 | 2023-09-12 | 王殿华 | AI algorithm system for highway construction safety management |
| CN117350501A (en) * | 2023-10-23 | 2024-01-05 | 青岛智能产业技术研究院 | System and method for dispatching man-in-loop power grid driven by large language model |
-
2024
- 2024-03-20 CN CN202410320595.2A patent/CN118214676B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114254649A (en) * | 2021-12-15 | 2022-03-29 | 科大讯飞股份有限公司 | A language model training method, device, storage medium and device |
| CN115186473A (en) * | 2022-07-07 | 2022-10-14 | 北京交通大学 | Scene perception modeling and verifying method based on parallel intelligence |
Also Published As
| Publication number | Publication date |
|---|---|
| CN118214676A (en) | 2024-06-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN118214676B (en) | Network fault control method based on parallel network large model digital expert | |
| Li et al. | A survey on LLM-based multi-agent systems: workflow, infrastructure, and challenges | |
| Bobillo et al. | Joining Gödel and Zadeh fuzzy logics in fuzzy description logics | |
| CN111737432A (en) | Automatic dialogue method and system based on joint training model | |
| CN119557409B (en) | Intelligent question answering method and system based on multi-module collaborative optimization | |
| Gao et al. | Modular rag: Transforming rag systems into lego-like reconfigurable frameworks | |
| CN117235216A (en) | Knowledge reasoning method based on heterogeneous knowledge fusion | |
| CN119475217B (en) | Key information demand intention understanding method and system based on large model | |
| CN115964459A (en) | Multi-hop reasoning question answering method and system based on food safety cognitive map | |
| CN118467854A (en) | An intelligent optimization system for low-code development platforms based on machine learning | |
| CN119294520A (en) | Multi-agent question answering system based on large language model | |
| CN113836944B (en) | Control method, system and device for smart device, and electronic device | |
| CN119917859A (en) | Network monitoring model training method, monitoring method, system, equipment and medium | |
| CN119692477A (en) | Hybrid question-answering method, device, equipment and medium for hybrid expert large language model | |
| CN118036750A (en) | Embodied intelligent task planner training method and system based on multimodal large model | |
| Cai et al. | Large language model-enhanced reinforcement learning for low-altitude economy networking | |
| CN117544508B (en) | Network equipment configuration query method and device, terminal equipment and storage medium | |
| Zhang et al. | Enhancing Logical Reasoning in Large Language Models via Multi-Stage Ensemble Architecture with Adaptive Attention and Decision Voting | |
| CN119961288B (en) | Intelligent recommendation method, intelligent control method and device for load transfer strategy of distribution network | |
| Yang et al. | Prospects and Challenges of Large Language Models in the Field of Intelligent Building | |
| CN118195032B (en) | Large model automatic evolution system and method with active learning capability | |
| KR102385198B1 (en) | Conversation generating system and method for chatting between artificial intelligences | |
| Zhao | Automatic search of machine learning models based on intelligent computing | |
| Ma | Dynamic Rule Extraction Method Based on Computer Fuzzy Chinese Language Recognition System | |
| CN120336496A (en) | Dynamically optimized large language model cluster collaborative question-answering method, system, device and medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |